Lunar surface geometry reconstruction via 3D Gaussian Splatting

Surface geometry reconstruction of lunar scenes is a key technique for understanding lunar terrain, which enhances perception and decision-making to enable efficient and autonomous lunar rover exploration. However, the unstructured and unbounded nature of lunar scenes substantially limits surface 3D reconstruction quality. To mitigate this limitation, we introduce a dedicated framework based on Gaussian Splatting for complete and accurate surface reconstruction in lunar environments. Our method features two key innovations. First, a progressive spatial-frequency joint supervision strategy is proposed, in which frequency-domain cues from input images and spatial constraints derived from Gaussian splats are integrated to guide scene reconstruction in a coarse-to-fine manner. Second, an optimization strategy guided by geometric priors is proposed, effectively improving scene representation. Experiments on authentic lunar surface scenes demonstrate that our method is capable of reconstructing the accurate and complete 3D mesh model, with additional experiments on novel view synthesis further validating its effectiveness from visual fidelity.